Watch Me Build an AI Chat Agent Solution for a Real Client

Key Takeaways

in this video you'll watch me build an AI chat agent for a client the client is an ebike rental business in Berlin they want a chat feature on their website capable of answering all customer questions for example how much an ebike costs which models are available and if there's a nearby location they provided files containing all the necessary information in the next few minutes I'll demonstrate how you could handle similar situations and build the same chatbot yourself I'll show you a straightforward method using n8n a no code tool that allows us to create the necessary workflows later I'll demonstrate how to use it to easily populate your own Vector store with the required data and build a chat that accesses this data I'll also provide example code for accessing a provided chat endpoint to integrate the chat into your web applications the complete process consists of three steps prepare the data given by the client build a workflow that inserts the data into a vector store build a chat agent that can access the vector store to instantly answer based on this data this video will equip you with all the knowledge needed to build these types of chat Bots for yourself or to offer as a service to clients if you're considering starting your own AI side hustle whether through AI SAS applications or freelancing visit AI for dev.com without further Ado let's dive in and build this solution together after completing the free registration we arrive at the main interface where we can immediately begin creating a workspace our initial step is to add a manual trigger next we'll establish a connection to Google Drive we search for Google Drive and click on the first result in this step we want to search files within the drive we first need to establish a connection to the Google Drive API to create this connection we need a client ID and a client secret we'll obtain these from the Google Cloud console first ensure the Google Drive API is activated indicated by a check mark if not activate it by clicking the appropriate button our next step is to create the client ID and client secret navigate to credentials select create credentials and choose o or client ID select web application assign a name and input the URL from n8n after creation you'll receive a client ID and client secret enter these in n8n and save test The Connection by clicking sign in with Google and confirming the dialogue a successful connection will be indicated now let's add a filter to select only files from the Berlin bike folder test this step to confirm it recognizes all relevant text files now that we have access to the files we need to download them to the workspace before adding them to the vector store for file downloading search for Google Drive again and select download file choose the credentials we just created we'll use the ID from the current step for each file and replace the file name with the actual name test this step to ensure proper file downloading and it looks good we see that the files have been downloaded accordingly we can now take care of the integration within pine cone for this we search for pine cone and select the pine cone Vector store we see that an API key is needed to connect to access our pine cone instance we go to the pine cone page and create a so-called Index this corresponds to a database in a relational database we select text embedding three small as the embedding and since we want to create a free index we use Amazon as a provider we see that an empty index is now being created and we can use it right away we then click on API keys and create a new API key we copy this and can now enter it here this worked well too we can save the files in the new index bike Berlin now we have to tell it how the embeddings should be created and here we select the same embedding again but in order to make it possible we first have to connect it to open AI That's why I copy my open AI key and now select the same embedding model again text embedding three small we now have to adjust at the vector store that we want to add data for this we click on Pine Cone Vector store and select insert documents as operation mode when we have done that we see a new plus symbol and here we can then select how the data is loaded we select binary and leave all other options exactly as before we have to tell it how large texts should be split and for this we take the simplest the token splitter which should split the texts all 5,000 tokens and there should be an overlap of 1,000 tokens good then we can execute it for this we just click on the small play button okay that looks very good and we can now look in Pine Cone whether the data has been correctly entered in our index we see several entries and if we go to the clear text view then we see the content of the different text files with information such as the locations or the bike types perfect in the next step we will now take care of building a chat that can respond to this data we start here with a trigger that starts when a chat message is received and we add an AI agent node in the next step we choose the option conversational agent and give him a system message that fits his task better we tell him you're a helpful AI assistant that helps bikes Berlin to provide useful answers to customers questions you're humorous and you don't make things up if you don't find the answer ask the customer to call the hotline now we have to assign a language model to the agent we use GPT 4 for this as memory we use a simple window buffer memory which keeps the last five messages local like a kind of short-term memory we tell him that he can call a vector store via the vector store tool and as a description of this tool we say that he should call this tool if he wants to give relevant and useful answers to the customer all right let's tidy up a bit to keep everything organized now we have to select that the database under the tool is our pine cone database and we enter the corresponding index now we need to tell him which embedding was used initially we can do this by clicking on the small plus icon labeled embedding and to correctly interpret the embedding we select text embedding three small again then we have to determine a language model that interprets the results of pine cone here we use again gp40 I think our Masterpiece is complete and we have a great first version we have an agent that is triggered whenever a new message is received the chat has short-term memory and can access a tool connected to the pine cone database where our documents are already embedded now we can try out our chat how much does an ebike cost for one hour the calculation looks good it says that an ebike costs €1 based on the data in the text document I ask again if he has bikes near Alexander Plotz this question is also correctly answered from the information from the text documents now we have a question that is a bit more complex and he says he can't answer it but he correctly refers to the hotline perfect now I want to show you how to install this chat on a real website with custom code we see here that we have a chat URL that can be reached from outside we can use this URL in our code to address the chat I have here a small example script built which uses this URL and a small web UI you can find a link to the code in the description it's a very basic JavaScript node project that gives you a starting point to integrate these chats in your websites